In this contribution, we present a computational stylistic study and comparison of classic French literary texts based on a datadriven approach where discovering interesting linguistic patterns is done without any prior knowledge. We propose an objective measure capable of capturing and extracting meaningful stylistic syntactic patterns from a given author's work. Our hypothesis is based on the fact that the most relevant syntactic patterns should significantly reflect the author's stylistic choice and thus they should exhibit some kind of peculiar overrepresentation behavior controlled by the author's purpose with respect to a linguistic norm. The analyzed results show the effectiveness in extracting interesting syntactic patterns from novels, and seem particularly promising for the analysis of such particular texts.
Computational stylistics is a subdomain of computational linguistics located
at the intersection of several research areas such as natural language
processing, literary analysis and data mining. The goal of computational stylistics
is to extract style patterns characterizing a particular type of texts using
computational and automatic methods
However, the process of extracting textual patterns is known by its property of producing a large amount of patterns, even from a relatively small sample of text. Thus, a measure of interest is to be applied to identify the most important and relevant patterns for the characterization of the text’s style in question.
In this paper, we present a computational stylistic study of classic texts of French literature based on a data-driven approach where the discovery of interesting linguistic forms is done without any prior knowledge. Specifically, the proposed method is based on the assessment of the peculiar overrepresentation of syntactic patterns extracted using sequential data mining technique from texts with respect to a norm corpus. This method is intended to quantitatively support a textual analysis by focusing on the verification of the degree of importance of each syntactic pattern (syntagmatic segments with potential gaps), and by extracting the syntactic patterns that characterize the syntactical style of a work by a particular author. 2
Approach for extracting relevant syntactic patterns Our method consists of two steps. First, a sequential pattern mining algorithm is applied to the texts in order to extract recurrent syntactic patterns. Second, a peculiarity-based interestingness measure that evaluates of the overrepresentation (in terms of frequency of occurrence with respect to a norm corpus) is applied to the set of extracted syntactic patterns. Thus, each syntactic pattern will be assigned an interestingness value indicating its importance and its relevance for the characterization of text’s syntactic style. In what follows, we present in section 2.1 the corpus used in our experience, and its dividing protocol into two parts: text to analyze and text used as norm. Then, section 2.2 introduces some elements necessary to understand the process of extracting sequential syntactic patterns. Finally, the formulation and the statistical details of the proposed interestingness measure are presented in Section 2.3.
In our study, we used four novels, belonging to the same genre and the same literary time span, written by four famous classic French authors: Balzac’s “Eugenie Grandet”, Flaubert's “Madame Bovary”, Hugo’s “Notre Dame de Paris” and Zola’s “Le ventre de Paris”. This choice is motivated by our particular interest in studying the style of the classical French literature of the 19th century. At the time of the analysis of the syntactic patterns, each text written by one of the four authors is contrasted with texts written by the three other authors. That is to say that these three texts will be considered as norm corpus from which we will evaluate the hypothesis of the overrepresentation of syntactic patterns in the fourth remaining text, as explained later in this section. 2.2
In our study we consider a syntagmatic approach. The text is first segmented
into a set of sentences, each sentence is then represented by a sequence of
syntactic labels (POS-tag)1 corresponding to the words of the sentence using
Treetagger
To avoid the effect of statistical fluctuations on the analysis of patterns with low supports, we considered a support’s threshold of 1%. That is to say that we focus only on patterns that are present in at least 1% of the sentences of the analyzed text. However, as sequential pattern mining is known to produce a large quantity of patterns even from relatively small samples of texts, 1 Frech treetagger tagset:
http://www.cis.unimuenchen.de/~schmid/tools/TreeTagger/data/french-tagset.html 2 <*> denotes a gap that can be filled with any POS tag an interestingness measure should be applied on these patterns in order to identify the most important ones. This interestingness measure is explained in the next section. 2.3
Evaluation of the relevance of syntactic patterns Our hypothesis to evaluate the relevance of a syntactic pattern is based on the fact that the most relevant ones should significantly reflect the stylistic choice of the author and should thus be characterized by a significant peculiar quantitative behavior, this peculiar behavior translate into a support’s overrepresentation in his texts.
However, to capture this overrepresentation one cannot refer only to the absolute frequency of occurrence (support) Indeed, more frequent use of a syntactic pattern by an author (which translates into a relatively high support) does not necessarily indicate a stylistic choice since it can be very well a property imposed by the grammar of the language or by syntactic features that are characteristic of text’s genre.
Thus, to assess the over-representation of a pattern, we use an empirical approach based on the comparison of the support of a syntactic pattern in a text to that found in a norm corpus. A ratio 4 between these two quantities is calculated as follow: 4 = frequency of a pattern in the norm corpus
frequency pattern in the text
In our experiments we found empirically that the distribution of the ratio 4
exhibits a Gaussian behavior. Indeed, the values of the 4 ratio are normally
distributed around a central value (see Fig. 1). This is due to the fact that the
frequency of occurrence of a syntactic pattern in a text is highly correlated
with the frequency of occurrence in the norm corpus with a few exceptional
special cases or outliers (see Fig. 2). These outliers represent the patterns of
special interest for our study because they represent a certain linguistic
deviation that is specific to the author's style compared to what one would expect
to see in the norm corpus.
The configuration described above allows us to use an outlier detection
method based on Gaussian distribution and '-score to identify such special
patterns
Results and Discussion
In this section, we present some examples of relevant syntactic patterns
extracted from our corpus. Using the proposed method, the extracted patterns
seem to have a strong relevance to characterize the style of the authors of our
corpus but also to the novels’ content and the literary genre in which it
operates. In the Flaubert's Madame Bovary, several extracted patterns well
represent the rhythmic rather than functional role of punctuation that is peculiar
to the style of Flaubert
Pattern (1) <PUN> < KON>< PUN> <PRP>, with support= 113, sample instances of the pattern in the text: • , et , à • , mais , avant • ; et , à In le Ventre de Paris of Zola, and in the same direction, the syntactic patterns extracted as relevant clearly represent the use of nested clauses to describe situations or attitudes in the novel such as in the pattern (2), or to describe public places and objects in displays in long lists as in the pattern (3):
Pattern (2) : <PUN> <PRP> <PRP> <NOM>, support= 104, sample instances of the pattern in the text (bold text): « Florent se heurtait à mille obstacles , à des porteurs qui se chargeaient , à des marchandes qui discutaient de leurs voix rudes ; il glissait sur le lit épais d' épluchures et de trognons qui couvrait la chaussée , il étouffait dans l' odeur puissante des feuilles écrasées .»
Pattern (3): <NOM> <PUN> <PRP> <NOM> <ADJ>, support= 68, sample instances of the pattern in the text (bold text): • angles , à fenêtres étroites • très-jolies , des légendes miraculeuses • écrevisses , des nappes mouvantes
In Eugénie Grandet of Balzac, other different communicative functions are performed by the syntactic patterns and their textual instances, for example:
Pattern (4): <PUN> <VER> <NAM> <PRP>, support= 49, which is used as post-introducer of direct speech. This rather formulaic way of specifying (in a parenthetical form) the utterer of a reported speech is common to all, but seems to be strongly preferred by Balzac, while the other authors have shown a more varied style in introducing dialogues. Sample instances of the pattern in the novel: • , dit Grandet en • , reprit Charles en • , dit Cruchot en
Pattern (5): <NUM> <NUM> <NOM>, support= 54, is a pattern used to refer to money, which is typical for the novel scenario where money plays a very important role. Sample instances of the pattern in the novel: • vingt mille francs • deux mille louis • sept mille livres
Pattern (6) : <ADV> <VER> <PRO> <ADV>, support= 59, is used to express negative questions : • n' avait -il pas • ne disait -on pas • ne serait -il pas
Pattern (7) : <PUN> <NOM> <PUN> <VER>, support= 44, represent the punctuation extensively used to mimic spoken intonation and even to reproduce performance phenomena such as stutter. : • , messieurs , cria • , madame , répondit • , mademoiselle , disait
The few analyzed examples indicate that the presented technique is effective in extracting interesting syntactic patterns from a single text, and this seems particularly promising for the analyses of such classic literary texts. On the other hand, this technique, as well as other similar ones, prompts the question of what is really captured by significant patterns. Some structures may be significant because they are typical of an author’s style, its fingerprint - as we may say borrowing a metaphor often used in attribution studies, or they may be dictated by functional needs, due to the particular topic of the novel, or to the conventions of the chosen genre. This is particularly true for syntactic analysis, where the functional constraints on the authorial freedom are more evident. Much further works have to be carried out concerning this issue. 4
conclusion In this paper, we have presented an objective interestingness measure to extract meaningful stylistic syntactic patterns from a given author’s work. Our hypothesis is based on the fact that the most relevant syntactic patterns should significantly reflect the author’s stylistic choice and thus they should exhibit some kind of peculiar overrepresentation behavior controlled by the author’s purpose. To evaluate the effectiveness of the proposed method, we conducted an experiment on a classic French Corpus. The analyzed results show the effectiveness in extracting interesting syntactic patterns from this type of text.
Based on the current study, we have identified several future research directions such as exploring other statistical measures to assess the interestingness of a given syntactic pattern, and expanding the analysis to include morphosyntactic patterns (form and lemma words). Finally, we intend to experiment with other languages and text sizes using standard corpora employed in the field of computational stylistics at large.